Concept
Project Concept: To train Stable Diffusion 1.5 to generate images in the style of 1920s Soviet avant-garde posters, using this style to tell stories about contemporary life. Stylistic reference: Constructivism—flat geometric forms, rigid diagonal composition, a palette of red, black, and cream, photomontage presentation, and the texture of old printing.
The dataset was compiled from the base model’s own generations—a style anchor plus 30 variations obtained via txt2img and img2img with varying stylization strengths; this is the cleanest option from a legal standpoint, as no external works were used.
Generations
I trained DreamBooth LoRA rank 16 on the trigger token ohwx for 800 steps. I was interested in testing whether the style held up on subjects far removed from 1920s aesthetics and to see where the adapter would fail.
1 — a food delivery courier with a thermal backpack, 2 — an office worker at a laptop computer
From the very first generations, it becomes noticeable how the generations adapt to modern realities. Recognizable elements of the posters remain, and it is clearly evident that models have the most difficulty transforming contemporary objects.
1 — a barista behind an espresso machine, 2 — passengers waiting on a subway platform
Despite the absence of a timeframe description, the neural network generates images oriented around the 50–60s era.
A doctor with a stethoscope in a hospital ward
A distinguishing feature is the ability to render landscape details and large-scale objects exceptionally well, even without excessive detail. It is also apparent that the color palette can sometimes diverge from the flat style of the initial generations.
My personal favorites from the series are the posters featuring plants and the car. The images turned out quite detailed, yet they still managed to retain the poster motifs of their references. The color palette of both images deviates from the typical look of posters, which made them more engaging.
1 — A cyclist crossing a bridge in the city, 2 — A gardener watering potted houseplants on a balcony
1 — A football striker kicking the ball. 2 — An electric car plugged into a charging station.
1 — a rock guitarist on stage in front of a crowd, 2 — students at desks in a university lecture hall
Overall, the style translates confidently: flat shapes, diagonals, and pseudo-typography are present, and the weakest elements are the natural subjects, to which constructivism is historically unsuited. The project demonstrates how to capture a unique 'stylistic voice' from a neural network in a single evening and what compromises emerge in the process.
a skateboarder jumping down stairs
Generative Model Application Description
In my project, generative AI was employed in three roles: Stable Diffusion 1.5 was used to create the training dataset and establish a stylistic anchor, and subsequently, after fine-tuning, it was used to generate the final series of posters. YandexGPT 5 was utilized for supplementary project texts—drafting captions for the dataset images, sketching the idea description, and reviewing the series—with the language model only drawing from factual data and numbers sourced from real runs.
